Análise Preditiva do Nível de Água no Canal São Gonçalo Utilizando Modelos de Aprendizado de Máquina: uma Abordagem Integrativa para a Gestão Hídrica

Abstract

In the field of hydrology, artificial intelligence models have stood out as effective tools in various research efforts. Accurate prediction of water levels in watersheds and rivers plays a crucial role in flood prevention strategies, inland navigation manage ment, and ensuring domestic water supply. However, a comprehensive analysis of the applicability of these models, especially in the context of predicting water levels in vast datasets, has been scantily explored. This study investigates and compares the performance of various artificial intelli gence models, including dense neural networks, recurrent neural networks, random forest, and support vector regression, in predicting water levels. Additionally, hybrid architectures are proposed to optimize predictive accuracy. The practical evaluation of these models is conducted through a case study in the São Gonçalo Channel. The study encompasses four distinct analyses, each based on datasets developed in col laboration with specialized water engineers. Beyond the comparative evaluation of artificial intelligence models in water level prediction, this work presents an innovative contribution by introducing hybrid architec tures that combine the potential of the ARIMA model with traditional machine learning approaches. The integration of these methods aims to capitalize on the intrinsic ad vantages of both, promoting a combinatorial approach to optimize the accuracy and robustness of hydrological predictions. This hybrid approach reflects a less-explored perspective in the development of advanced strategies for anticipating water levels, suggesting new possibilities to enhance the efficiency of water management systems. The study thus highlights an innovative dimension in the field, contributing to the ad vancement of knowledge and practices in hydrological forecasting. Another significant contribution of this thesis was the meticulous collection and anal ysis of climatological and water level data, serving as the basis for the development of four distinct datasets, each characterized by its specificities. This effort not only ex panded the understanding of the addressed climatological scenario but also facilitated the creation of representative and diversified datasets. The careful selection and ma nipulation of these datasets allowed for a more comprehensive and in-depth approach to predictive modeling, enriching the research with valuable insights into the complex dynamics between climatic factors and water levels. This multidimensional approach strengthens the theoretical and practical foundation of the thesis, contributing to the robustness and relevance of the obtained results. The results obtained reveal that hybrid models demonstrate superior predictive per formance compared to isolated approaches. This finding reinforces the viability and effectiveness of machine learning techniques in the hydrological sphere, highlighting their potential as decision support tools in contexts related to water resource manage ment.